The Complete Guide to Where AI Fits in a $5M to $100M Company, Department by Department

Every founder I meet at this scale has already bought the licenses. ChatGPT seats for the office, an AI notetaker on the sales calls, one pilot project a department head championed for a quarter. Ask what AI does for the business and you hear "everyone uses it." Ask which department owns a workflow that AI runs end to end, with a number attached, and the room goes quiet.

I watch this from two seats. I run AI and growth at NuVision Auto Glass, a $48M US auto glass company, and I founded NuroSparx, a growth company serving US brands between $5M and $100M. The pattern repeats everywhere. Companies rolled out AI as a personal productivity perk, so it made a hundred small tasks slightly faster and changed nothing you can see on the P&L, the profit and loss statement that tells you whether the quarter worked.

You'd never make a hire that way. Every hire gets a department, a job description, a manager, and a number they answer to. Treat AI the same way. Walk your org chart one department at a time. In each one, give AI a specific job, mark what it only drafts, and mark what it never touches.

Start with sales, because the math there is the easiest to check.

Sales: the work between conversations

You run reps with quotas, a CRM, which is the system that tracks every deal and conversation, and a weekly pipeline review. AI belongs in the gaps between those conversations, because that's where deals die.

The jobs it owns: writing call notes into the CRM minutes after the rep hangs up, researching an account before the discovery call, drafting the first version of a proposal from the call transcript, and running the follow-up sequence after that proposal goes out.

Follow-up comes first. A proposal ships, the rep chases the hotter deal, and the third and fourth touches never happen because they're nobody's job on a Tuesday afternoon. That's exactly the shape of work AI handles well: repetitive, text based, time triggered. Three days after the proposal, a short note referencing the specific scope. Seven days later, a different angle. Fourteen days, a final one. The rep approves every message before it sends, which costs minutes a day instead of the hour it takes to write them.

Run the check on your own pipeline. Count the proposals your team sent last quarter. Count how many received a third touch. If the second number sits under half the first, sales is your first department.

Keep pricing, negotiation, and closing with your reps. AI prepares the ground; humans play the deciding points.

Marketing: volume without losing the voice

This is where most teams start, and the trap here is bigger for you than for a startup, because you have a brand with revenue behind it.

The useful jobs: turning one flagship asset into every downstream format, drafting landing pages and email sequences, replying to every review within a day, and mining your own sales calls for content. Teams that use AI in content report production costs down 40 to 60%, and a lot of that output reads identically to everyone else's. At your scale, publishing same-sounding AI drafts at volume trades a decade of brand voice for throughput. The edit pass is the budget line that protects it.

One rule for this department. AI writes the draft, and someone who has sat in your sales calls writes the specifics. The details that make a buyer believe you live in your team's heads, not in the model's training data.

A concrete version: have your team feed ten won-deal call transcripts into the model and ask for every objection that shows up in at least three of them. That list becomes next quarter's content calendar, built from what your buyers actually said. No competitor can copy it, because they don't have your calls.

One more job now sits in this department whether you assigned it or not. Your buyers ask AI assistants who to shortlist before they ever open a search page. Marketing owns whether your company appears in those answers, the same way it owned your Google rankings for the last twenty years.

Operations: get the business out of people's heads

Every company at this scale runs on knowledge that lives in senior people's heads and nowhere else. New hires learn by shadowing. Quality varies by who trained whom. One resignation walks a decade of judgment out the door.

AI removes most of the pain of fixing that. The protocol: each department head records a 10 minute voice walkthrough of one core process, exactly the way they want it done. Transcribe it. Hand the transcript to the model and ask for a step-by-step guide a new hire can follow. The manager corrects the steps the model guessed at, once. Repeat across the 10 most common processes in each department and you've built the operations manual every company your size claims it will write eventually.

On the ground, it looks like this. An ops manager records herself explaining how to close out a job, including what to photograph and what to tell the customer before leaving. Eleven minutes of talking, most of it rambling. The transcript goes in, a checklist comes out, and she edits the three steps the model guessed wrong. About 25 minutes total for a document she'd meant to write for years, and every new hire now closes jobs the same way.

I see the stakes daily at NuVision, a $48M service operation where a week's outcome rides on scheduling, dispatch, and handoffs between teams. A documented handoff survives a resignation. An undocumented one is a person, and people leave.

Finance: AI prepares, a human approves

Your finance team closes the books, chases receivables, the money customers owe you, and reviews contracts. AI stays on the preparation side of all three.

Jobs that work: turning invoices and receipts into structured rows, drafting the collection email for anything past 30 days, summarizing the month's numbers in plain English for the leadership meeting, sorting expenses into categories on a first pass, and reading supplier contracts to pull out dates, penalties, and renewal terms.

Jobs that don't: filings, final tax numbers, and any figure that reaches a government body or a bank without a human checking it. The model produces a confident, correctly formatted, wrong number as easily as a right one, and it will not flag which is which.

A concrete version: renewal terms. Your controller drops the year's supplier contracts into the model and asks for one table: vendor, renewal date, notice window, penalty. Auto-renewals stop surprising the P&L, because every notice window now sits on a calendar instead of inside a PDF nobody reopened.

The split holds every time. AI prepares, a human approves.

Support: triage, drafts, and the night shift

At your scale, support is a team with a queue, response-time targets, and a manager who reports on both. AI fits in three places.

1. Triage, which means sorting tickets by topic and urgency. The model reads every incoming ticket, tags it, routes it to the right person, and flags the customers who sound ready to walk.

2. First drafts. The model drafts an answer from your documentation and the customer's history before the agent even opens the ticket. The agent edits and sends, and stays accountable for every word.

3. The night shift. AI answers routine questions from your documentation after hours and books callbacks, with a clean handoff to a human every morning.

Picture the warranty question that lands at 9pm. Today it sits in the queue overnight while the customer stews. With this in place, the customer gets the documented answer straight away plus a morning callback on the calendar, and your team walks into a sorted queue instead of a pile.

What stays human: angry customers, refunds, and anything with legal weight. The pattern behind that list applies in every department: high cost when wrong, low ability to catch the error before it lands.

How to pick the first department

Score each department on two things: how often the work repeats and how much payroll it eats. Multiply the two. Start with the highest number.

Then hold three rules:

1. One department at a time, for 30 days. One working workflow beats six half-built ones, and your team's patience for AI projects is a budget you spend once.

2. Name the number before you start. Proposals with no third touch. First-response time on tickets. Days from invoice to cash. A workflow without a number attached is a demo.

3. Judge at day 30 and no earlier. Most deployments show a measurable change inside 2 to 4 weeks. If nothing moved by day 30, you picked the wrong department, and that lesson cost you one month instead of a year.

The companies pulling ahead at this scale didn't buy better tools. They gave AI a desk, a job description, and a manager, then left it in place long enough to measure. Everyone else bought licenses and a vague feeling of progress.

Pick one department this week. Name one number. Check it in 30 days.

If you want help picking the department and building the first workflow so it holds up under real volume, that's the work we do at NuroSparx. Tell us which department is costing you most and we'll map it with you. You can see what the finished version looks like in our case studies. Or book a call, walk us through your org chart, and we'll tell you which department we'd start with. The briefs your team will need on day one live in the first piece of this series.